Skip to content

Data Engineer interviews

How to interview a Data Engineer in 2026

Give them a small version of the real job: source data with the mess real systems make, the tables the business needs and an independent figure those tables must match. Let them use an AI assistant to write code, then judge what the AI cannot do for them: find what is wrong in the data, prove the tables are right and explain the call. A BuildInterview Audition does this in under two hours, and code checks the work one problem at a time.

Time
100 to 115 minutes
Levels
Junior, mid, senior and staff
Live today
Build Auditions in e-commerce, junior to staff. Fix Auditions in banking (mid to staff), e-commerce (mid), SaaS (senior), media (senior) and logistics (staff).
Tools
Python and SQL, plus dbt, Apache Airflow and a system design canvas when the question needs them
Price
Pay per candidate, no subscription. Every approved company gets up to 5 interviews free.

Updated

What does a data engineer do in 2026?

A data engineer builds and runs the pipelines that move data from the systems where it is created into tables other people rely on: finance, operations, analysts and, more and more, AI assistants and agents. The tables have to be right, on time and safe to rebuild when something goes wrong.

The core skills have not changed. When we wrote our Data Engineer Auditions in October 2026, we read 21 open job postings at four levels across eight sectors. SQL and Spark each appeared in 16 of them and Python in 15. Of the 21, 18 asked for data quality or testing and 19 for work with the people who use the data.

What has changed is who writes the code. AI assistants now write much of the SQL and Python, so the job leans harder on checking: knowing what the data means, proving a table is right and owning it when it is not. That is the part an interview has to show.

What should a data engineer interview test?

Six things decide whether a data engineer’s tables can be trusted. A good interview lets you see each one in the work, not just hear about it.

  1. Reading the data first. Do they look at the files, the keys and the data dictionary before writing code, or start from the first line of the brief?
  2. Getting the grain and the rules right. Does each table count what it says it counts, one row per order, per day or per customer, with every published rule applied in order?
  3. Making it safe to run again. If the same batch arrives twice, or a day is replayed, does every published table stay exactly the same?
  4. Proving the tables are right. Do they reconcile what they publish with an independent source of truth, and account for every record they set aside?
  5. Checking what the AI wrote. When the assistant writes a query, do they test it on the cases that break pipelines, such as repeats, late records and missing values, or trust it because it runs?
  6. Making the call. Can they say whether the data can be used, explain it to the people who rely on it and leave a runbook the next engineer can follow?

Which interview format shows the job?

Most data engineering interviews use one or more of four formats, and each shows something different.

  • Algorithm tests. Quick to run for many applicants, but they test puzzles a data engineer rarely solves at work, and an AI assistant can solve many of them.
  • Take-home tasks. Closer to the job, but they cost candidates hours of unpaid time, are hard to compare and do not show who or what did the work.
  • Live pair coding. Shows how someone reasons out loud, but an hour is too short for a pipeline on real data, and the result depends on the interviewer.
  • Work samples. A timed, realistic piece of the job, checked the same way for every candidate.

A data engineering work sample needs two things a coding test lacks. The first is real mess: the repeats, late records, broken rules and changed formats that real systems produce, so the candidate has something to find. The second is a source of truth: an independent figure, such as a control total or a statement, that the published tables must match. With both, you can tell a table that looks right from one that is right, whoever wrote the code.

Work samples also hold up in hiring research. In the largest recent review (Sackett and colleagues, Journal of Applied Psychology, 2022), work-sample tests predicted job performance better than general ability tests (0.33 against 0.31) and far better than unstructured interviews (0.19). Structured interviews came first (0.42), so the strongest loop pairs a work sample with a structured conversation about it.

For every format side by side, with what each misses when candidates use AI, read technical interview formats compared.

How the Data Engineer Audition works

An Audition is BuildInterview’s work sample. The candidate gets a brief from the data’s owner, the source files and a source of truth, and works in a browser workspace in three timed steps: Plan, Build and Writeup.

  1. Read the brief

    The data’s owner brings a business problem and a deadline. The source files carry the mess real systems make, and a source of truth says what the totals must be.

  2. Plan, with AI off15 minutes

    They write how they will find what is wrong in the data, what they will publish and how they will prove it. The plan locks and stays on screen, so you see their own thinking before any AI help.

  3. Build, with an AI assistant70 to 80 minutes

    They work in Python and SQL, plus dbt, Apache Airflow and a system design canvas when the question needs them. The AI assistant works like a pair programmer: it helps one step at a time, never plans the work and never writes the whole solution.

  4. Find the problems and prove the tables

    Problems are planted in the data, and each one shows up at least three times. They set aside what breaks the rules with a reason, make the load safe to run again and reconcile what they publish with the source of truth.

  5. Handle the update

    Unless you switch it off, the data’s owner changes something midway, as a real stakeholder would. The report shows when it arrived and what they changed.

  6. Write it up, with AI on15 to 20 minutes

    Two short notes: one for the people who use the data, with the call and its limits, and a runbook for the engineer who runs the pipeline next.

  7. After they submit

    Code runs their work again on test data they have never seen, one planted problem at a time, so answers typed in by hand fail. An AI reviewer rates what code cannot settle, and you get one report.

You can try the candidate’s workspace yourself in the demo, or read how evaluation works.

Build or Fix, and what the brief says

Every Data Engineer Audition is one of two types.

The two types of Data Engineer Audition
What the candidate doesLive today
Build a new pipelineBuild the load or the tables from the brief and the data.Live in e-commerce, from junior to staff.
Fix one that breaks the rulesTake over a job that runs but breaks the business rules, then find the problems and fix them.Live in banking (mid to staff), e-commerce (mid), SaaS (senior), media (senior) and logistics (staff).

No brief lists the problems. What it says about them depends on the level, so you see whether a candidate can find a problem nobody named:

  • Junior. A symptom for every planted problem, and every rule written down in order. A junior never designs the pipeline.
  • Mid. Symptoms for half of the problems and nothing about the rest, plus two or three published options to choose between.
  • Senior and Staff. Only the business problem, in the words of the person who has it. The brief gives the policy and the definitions, never how to handle the mess, so they find the problems themselves.

What changes from junior to staff

Titles and years do not set the level. One employer in our research posted two roles with the same title in the same week, one asking for 1 to 2 years and the other for 3 to 5. What sets the level is scope and judgment: how much is decided for the engineer, how large a thing they own and who follows their standards. The Auditions follow the same rule. More rows or a rarer tool never make a question harder; more judgment does.

Data Engineer Auditions by level
JuniorMidSeniorStaff
Experience0 to 2 years3 to 5 years5 to 8 years8+ years
The data’s owner asks forOne load or one table, with every rule givenA pipeline they own from source to table, where two or three rules are options they chooseA pipeline, a history or a repair for two users who want different things, within a deadline, a capacity or a budgetA decision across several pipelines or systems, backed by evidence they build
The brief describesA symptom for every planted problemSymptoms for half of themOnly the business problemOnly the business problem
Left to their judgmentNothing but the data’s messWhich published option to use, and showing its cost in the dataFinding the problems, the scope and how to weigh the two usersFinding the problems, and what to commit to, refuse or prove first
The source of truthA control total the published numbers must matchThe same, for every runThe same, and most planted problems break itSources that disagree, or an older system’s output
Kinds of planted data problem4 or 55 or 65 to 7, some acting on the same records5 to 7, plus sources that disagree
Time100 minutes110 minutes115 minutes115 minutes

What a strong candidate shows at each level:

Junior
Reads the data dictionary before writing code, applies every rule in the published order, counts at the right grain, runs the code and checks the results instead of trusting them, makes the reconciliation add up exactly, shows that a replayed batch changes nothing, and can explain every line, their own or the assistant’s.
Mid
Owns the flow from the landing file to the published table, makes the load incremental and safe to rerun, picks an option and shows its cost in the data, adds checks that would catch the failure the story describes, and leaves a runbook another engineer can follow.
Senior
Finds where late data, history and recovery act on the same records, repairs published data without breaking what was right, plans within the stated limit, and explains the trade-off to both users in their own terms.
Staff
Turns the evidence from several pipelines into one decision, builds a check others can reuse, separates what is proven from what is assumed, sets the rule others follow, and says what would change the call.

How the work is scored

Each Audition is scored out of 100 across the plan, the build and the write-up, with most of the weight on the build (70% by default). The build is scored issue by issue.

  • An issue is one real problem the work must solve: a problem planted in the data, the call the brief asks for, or the update when you send it. Every issue gets the same share of the build score.
  • Code checks each issue on two sets of test data the candidate never sees: one made from their own data with only that problem in it, and one fresh set. An issue is solved when every output is right on both. Only code gives an issue full marks.
  • An issue that code did not prove solved goes to an AI reviewer, which places the work on written levels. It never sees the expected answers, can never give full marks and never changes a solved issue. Every review runs twice, and if the two runs differ by two levels or more, a third run settles it.
  • The plan and the two notes are rated against written levels too. Hard-coded answers fail the fresh data, and missing evidence is marked not scored, never zero.
  • Polish, speed and how much the candidate used AI are never scored. AI use is shown beside the score, so your team can read it.
  • There is no hire or reject recommendation. The report opens the build with how many issues were solved and links every finding to the work behind it. Your team decides.

Read how evaluation works, or watch the report at the end of each Data Engineer demo film.

Questions for the follow-up interview

Use these in the structured conversation after the work sample. Ask every candidate the same questions and compare the answers against the same notes.

Someone reran yesterday’s job by mistake. What should that do to the published tables?
A strong answer: Nothing. Every row has a key, the load merges or replaces by that key instead of appending, and the run log shows no rows added by the second run.
Records for a day you already published arrive three days late. What do you do with them?
A strong answer: Follow a rule agreed with the people who use the table, such as reprocessing a fixed window of days or publishing corrections separately, make every run apply it, and tell users before a past day changes.
Your totals do not match the finance team’s numbers. Where do you start?
A strong answer: With the definitions: the grain, the time zone, the currency and what counts as one sale. Then reconcile step by step from the source rows to the published table, counting what was set aside and why, until the gap is explained.
An AI assistant wrote a query that runs and returns numbers. How do you decide whether to trust it?
A strong answer: Check it against something independent: a control total, a hand count on a small sample and the cases that usually break, such as missing values, repeats and late rows. Running without errors proves nothing.
A dashboard everyone uses showed a wrong number for a week before anyone noticed. What would you change?
A strong answer: Checks that run before publishing and stop a bad load, such as row counts against the source, totals against an independent figure and allowed values, an owner who gets the alert, and a short note to users about what was wrong and what changed.
Two teams want the same table. One needs it early every morning, the other needs it complete. How do you decide?
A strong answer: Make the trade-off explicit: what each team loses with each option. Often that means an early provisional version and a final one, clearly labeled, agreed with both teams in writing.
How would you backfill two years of history without disturbing tonight’s run?
A strong answer: In batches by date, written somewhere separate first, checked against the source before it replaces anything, with the nightly job left running and the time and cost estimated before starting.
The pipeline’s compute bill doubled this month. Where do you look?
A strong answer: First at work that is recomputed for no reason, such as full reloads of data that did not change, then at the largest and most frequent queries, and show the saving in numbers before changing anything.
What goes in the runbook you leave for the next engineer?
A strong answer: How to run it, how to rerun one day, how to backfill and recover, what each check means when it fails, who uses the tables and who to tell when something breaks.

Watch the demos, then try it free

Four demo films, one for each level, each follow a demo Data Engineer Audition from the brief to the report your team reads. Watch the Data Engineer demos, or see what each level asks for on the homepage.

Every approved company gets up to 5 interviews free, as a pilot. After that you pay per candidate, with no subscription. The price of each interview depends on what you include: the Auditions and questions, the AI model candidates use and the token budget for the task.

Hiring forward-deployed engineers too? Read the FDE interview guide.

For candidates: how to prepare

Data engineering interviews increasingly look like the job itself. To prepare:

  • Look at the data before you write code: the keys, the repeats, the gaps, the time zones and what one row stands for.
  • Make every load safe to run twice, and show that it is.
  • Reconcile what you publish with an independent total, and account for every record you set aside.
  • Check AI-written SQL on the cases that break pipelines, then explain your call in a few plain lines.

You can practise realistic data and AI interview questions on Lets Data Science .

Common questions

How do you interview a data engineer?

Give them a small version of the real job: source data with real mess, the tables the business needs and an independent figure those tables must match. Let them build or repair the pipeline, prove the tables are right and explain whether the data can be used. Then hold a structured follow-up conversation about the choices they made.

How do you hire a data engineer?

Start with the tables they will own and who relies on them. Set the level by scope and judgment, not years: how much will be decided for them and how much they will decide. Then test the job itself with a timed work sample on messy data, follow up with structured questions about that work, and ask references how the candidate handled broken data in production.

Should candidates use AI in a data engineering interview?

If they will use AI at work, yes, but test what the AI cannot do for them. AI assistants now write much of the code data engineers ship, so the skill that separates people is checking it: knowing what the data means, finding what is wrong and proving the tables are right. In a BuildInterview Audition, AI is off while they plan and on while they build and write up, and the assistant never plans the work or writes the whole solution.

How is a data engineer different from a data analyst or an analytics engineer?

A data analyst answers business questions with data that already exists. An analytics engineer models clean, tested tables in the warehouse, often with dbt, for analysts to use. A data engineer builds and runs the pipelines that bring data in from the source systems, keeps them reliable run after run and owns the tables others build on. The roles overlap, and in a small team one person may do all three.

What should a junior data engineer be able to do?

Follow written rules against messy data and get exact results. A strong junior reads the data dictionary before writing code, counts at the right grain, checks results instead of trusting them, makes a reconciliation add up exactly, shows that loading a batch twice changes nothing and can explain every line, including the lines an AI assistant wrote. A junior is not expected to design the pipeline or choose between methods.

What should a senior or staff data engineer interview test?

Judgment, not faster coding. A senior engineer should find problems nobody described, where late data, history and recovery act on the same records, and weigh two users who want different things within a limit. A staff engineer should turn evidence from several pipelines or systems into one decision, separate what is proven from what is assumed and set a rule others will follow. In a BuildInterview Audition, Senior and Staff briefs give only the business problem, so the candidate finds the problems themselves.

How long should a data engineering take-home or work sample be?

Long enough to show the whole job and short enough to be fair: about two hours in one sitting, with a hard time limit. Take-homes with no limit favor people with free time and are hard to compare. By default, BuildInterview Data Engineer Auditions take 100 minutes at junior level, 110 at mid level and 115 at senior and staff level, including the plan and the write-up.

Can a candidate hard-code the answers or let the AI do the work?

Hard-coding fails, because code checks every issue on test data the candidate has never seen. The assistant helps one step at a time and never writes the whole solution, so the candidate still has to find the problems, decide what to do and prove the tables are right. The report shows how they used AI beside the score.

Does the AI decide who we hire?

No. Code decides what is right wherever it can, an AI reviewer rates what code cannot settle against written levels, and the report shows the evidence behind every score. There is no hire or reject recommendation. Your team decides.

How much does a Data Engineer interview cost?

You pay per candidate, with no subscription. The price depends on what you include: the Auditions and questions, the AI model candidates use and the token budget for the task. Every approved company gets up to 5 interviews free, as a pilot, with Claude Haiku 5.5 or GPT-6 Luna as the candidate's AI assistant. The price calculator on our homepage gives an estimate for the interview you build.

Which industries and tools do the Data Engineer Auditions cover?

Build Auditions are set in e-commerce, from junior to staff. Fix Auditions are set in banking (mid to staff), e-commerce (mid), SaaS (senior), media (senior) and logistics (staff). Candidates work in Python and SQL, plus dbt, Apache Airflow and a system design canvas when the question needs them. More industries are on the way.

Interview your next data engineer on the real job

Create a company account, pick a Build or Fix Audition for the level you are hiring, and send it to your first candidate.